CN114818822A - Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation - Google Patents
Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation Download PDFInfo
- Publication number
- CN114818822A CN114818822A CN202210516213.4A CN202210516213A CN114818822A CN 114818822 A CN114818822 A CN 114818822A CN 202210516213 A CN202210516213 A CN 202210516213A CN 114818822 A CN114818822 A CN 114818822A
- Authority
- CN
- China
- Prior art keywords
- matrix
- label
- sample
- mapping
- target domain
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation. The method collects electroencephalogram data from different testees, and extracts features as a sample matrix. And constructing two sub-models of a double-mapping domain adaptation model and a semi-supervised regression and iconic notation propagation combined model for combined optimization. In the optimization process, different variables are subjected to alternate iterative optimization, so that the feature sharing subspace is continuously optimized, the data distribution difference between the source domain and the target domain is reduced, a more accurate target domain label can be learned, and the difference between the two domains is further reduced. The method can reduce the distribution difference of the electroencephalogram data among the tested subjects, further inhibit the influence of the difference existing among the tested data characteristics on the emotion recognition process, and solve the problem of poor recognition effect caused by the individual difference among the tested subjects in the cross-tested emotion recognition field.
Description
Technical Field
The invention belongs to the technical field of electroencephalogram signal processing, and particularly relates to an electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation.
Background
In 1990, Salovey and Mayer proposed the concept of emotional intelligence, and thought that emotional intelligence is an important component of artificial intelligence in addition to logic intelligence, for machines, it is a key to realize machine emotional intelligence to automatically and accurately identify the emotional state of human, and the emotion is a state integrating human sense, thought and behavior, including not only the psychological reaction of human to external environment or self-stimulation, but also the physiological reaction accompanying the psychological reaction, and the emotional expression of human, such as expression, voice, etc., can be disguised and changed unconsciously due to the influence of factors such as external environment and self-cognition, but the electroencephalogram signals from central nervous system activity are directly related to the real emotion of human, and are not disguised, so that it is widely applied to the field of objective emotional identification, and has become a research hotspot in the field, meanwhile, with the continuous development of subjects such as computer technology, biological science, neuroscience and the like, the electroencephalogram signals play an increasingly important role in improving the accuracy and reliability of emotion state recognition in human-computer interaction.
Although the electroencephalogram can be objectively recognized for emotional states due to the non-disguise property, the electroencephalogram has an important problem, the individual difference exists, and the electroencephalogram between people and the electroencephalogram of one person at different time have distribution difference, which causes that the machine learning method based on the same distribution is difficult to generalize. Therefore, the concept of migration learning is proposed, the basic idea of which is to use knowledge of the auxiliary domain to facilitate the task of identification of the target domain. Common migratory learning methods can be generally classified into four categories: model-based, feature-based, sample-based, relationship-based. The feature-based method is the most widely used migration method, and the method aims to learn a shared feature representation and map target domain and source domain data into a shared subspace in combination with distribution measurement strategies, such as maximum mean difference and the like, so as to reduce the difference between the two domains and achieve a state of basically approximate distribution. The calculation of the condition distribution is related to the labels of the source domain and the target domain, so that if an accurate target domain label can be learned, a better projection matrix can be obtained, the distribution difference between the source domain and the target domain can be reduced better to obtain a more excellent target label, the model identification precision is improved, and the reliability of emotion human-computer interaction is ensured. The mapping method of the two-domain original data can be divided into single mapping and double mapping, theoretically, compared with single mapping, the double mapping considers the individual characteristics of the two domains in the mapping process, but in order to prevent the individual characteristics of the two domains from being excessively reflected in the shared subspace, certain constraint needs to be carried out on the mapping matrixes of the two domains, therefore, theoretically, the single mapping is considered to be included in the double mapping, most of the prior art adopts a single mapping mode, the beneficial individual characteristics of the domains in the mapping process are ignored, and only the common characteristics which can be used for migration are reserved.
Meanwhile, due to the individual differences of electroencephalogram signals, the feature representations of different tested electroencephalogram signals aiming at the same emotion also have differences, most of the prior art only simply carry out cross-tested migration recognition and cannot well meet the requirements of human-computer interaction on emotion recognition accuracy, therefore, a certain learning method needs to be combined to optimize the recognition process, graph-based label propagation is a more applicable mode, the effect of the graph-based label propagation is reflected in the prior art (CN113157094A), but when the constructed undirected graph is used for label propagation between two domains, the correlation between a source domain and a target domain is considered, the correlation of samples in the domains is also considered, and the correlation in a double-mapping environment possibly influences the final recognition result due to the retention of the individual features of the two domains.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides an electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation, which is implemented by aligning source domain data and target domain data in a shared subspace through a projection matrix P s ,P t A sample-label mapping matrix W of semi-supervised regression, a sample-feature bipartite graph correlation matrix S and a target domain emotion label F t Performing joint iterative optimization, and obtaining better alignment data through continuous iterative optimization feature sharing subspace so as to be better applied to learning of the sample-label mapping matrix W and the sample-feature incidence matrix S, thereby obtaining more accurate target domainEmotion label F t The method is applied to learning of the shared subspace.
Step 1, collecting electroencephalogram data of a plurality of testees under c different emotional states.
And 2, preprocessing and extracting features of the electroencephalogram data acquired in the step 1, wherein each sample matrix X consists of electroencephalogram features of a testee, and the label vector y is an emotion label corresponding to the electroencephalogram features in the sample matrix X. Selecting two sample matrixes X from different testees as source domain data X respectively s And target domain data X t 。
And 3, constructing an electroencephalogram migration emotion recognition model combining semi-supervised regression and icon label propagation, wherein the electroencephalogram migration emotion recognition model comprises a double mapping domain adaptation model and a semi-supervised regression and icon label propagation combination model.
Step 3.1, establishing a double mapping domain adaptive model:
wherein the content of the first and second substances,
respectively mapping matrixes for mapping source domain data and target domain data into a shared subspace, wherein a superscript T represents transposition, p represents a shared subspace dimension, and d represents an original sample dimension;
to augment the source domain data tag, Y s The code is a one-hot type code of the source domain label, and c is the category number;
in order to augment the target domain data tag,
representing a column vector with elements all 1, F t One-hot type coding for target domain tags, n s ,n t The number of samples of the source domain data and the target domain data respectively, n ═ n s +n t Representing the total number of samples;
for diagonal matrices, diagonal matrix N dom The kth diagonal element is the reciprocal of the number of data samples of the kth source domain or target domain, wherein k is 1, 2. Wherein
Is a central matrix, and the central matrix is a central matrix,
is a unit matrix which is formed by the following steps,
represents the square calculation of Frobenius norm, | | P s -P t || 2,1 Mathematically, the row sparsity constraint of the difference value of the mapping matrix of the two domains sharing subspace is realized, and the significance of the row sparsity constraint is to prevent the over-embodying of the individual characteristics of the two domains in the mapping process, | | · | computationally 2,1 Representing the calculation of a 21 norm.
Step 3.2, establishing a semi-supervised regression and icon label propagation combined model:
s.t.G≥0,G1＝1,F≥0,F1＝1,
wherein the content of the first and second substances,
is a sample-to-label mapping matrix,
a matrix of sample labels is represented that,
represents X T The average of the difference between PW and F,
representing a sample-feature bipartite graph correlation matrix,
representing a sample-feature correlation matrix, 0 n ,0 p Are respectively dimension of
The matrix of all zeros of (c) is,
a graph association matrix representing an initialization constructed based on current shared subspace data,
is a sample-feature correlation matrix initialized based on current shared subspace data,
a matrix of labels representing the samples and the features,
is a graph Laplace matrix, a diagonal matrix
The diagonal elements are the rows S of the matrix, Tr (-) represents the trace of solving the matrix, and the formula (2)
Is a standard formula of semi-supervised regression,
is a row sparsity constraint on the matrix W,
is a bipartite graph construction and label propagation process.
Step 3.3, combining the two models constructed in the steps 3.1 and 3.2, integrating the double mapping domain adaptation model and the semi-supervised regression and icon label propagation combined model into a unified frame for combined optimization, and calculating a target domain emotion label F t The optimization model is as follows:
Step 4, according to the optimization model established in the step 3.3, aligning the projection matrix P of the source domain data and the target domain data in the shared subspace s ,P t A sample-label mapping matrix W of semi-supervised regression, a sample-feature bipartite graph correlation matrix S and a target domain emotion label F t And performing joint iterative optimization.
Step 4.1, initialize the data label F of the target domain t And a characteristic label F d Has a value of
And constructing an initial shared subspace mapping matrix P in a dimension reduction mode s ,P t 。
Preferably, the shared subspace mapping matrix is constructed by Principal Component Analysis (PCA).
Step 4.2, fixing the target domain data label F t And a shared subspace mapping matrix P s ,P t And updating the sample-label mapping matrix W, wherein the objective function is as follows:
and solving the formula (4) to obtain an updated sample-label mapping matrix W'.
Step 4.3, fixing the sample-label mapping matrix W and the target domain data label F t Updating the shared subspace mapping matrix P s ,P t The objective function is:
solving the formula (5) to obtain the updated shared subspace mapping matrix P s ′,P t ′。
Step 4.4, fix the shared subspace mapping matrix P s ,P t Target domain data tag F t And a feature tag F d Updating the sample-feature bipartite graph correlation matrix S and calculating a new feature label F d The objective function is:
s.t.G≥0,G1＝1
solving the formula (6) to obtain an updated sample-feature bipartite graph incidence matrix S' and a feature label F d ′。
Step 4.5, fixing the sample-label mapping matrix W and sharing the subspace mapping matrix P s ,P t And sample-feature bipartite graph incidence matrix S and feature labels F d Update target Domain data tag F t The objective function is:
s.t.F≥0,F1＝1
solving the formula (7) to obtain an updated target domain data label F t ′。
Step 4.6, repeating steps 4.2, 4.3, 4.4 and 4.5 for multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P s ,P t Sample-feature bipartite graph correlation matrix S and target domain data label F t Joint iterative optimization of (2).
Step 5, substituting the result obtained after the combined iterative optimization in the step 4 into a formula (3), and then substituting the electroencephalogram data characteristic X with unknown emotional state into the electroencephalogram data characteristic X m As target domain data X t Inputting the data into a formula (3), and solving a corresponding target domain data label F t Obtaining the characteristics X of the brain electrical data m And correspondingly acquiring the emotional state at the moment.
The invention has the following beneficial effects:
1. the electroencephalogram migration emotion recognition model combining semi-supervised regression and icon label propagation provides an effective tool with higher accuracy for emotion man-machine interaction, and the target label is continuously optimized in an iterative mode through a mathematical model, so that the emotional state of a testee can be accurately recognized according to electroencephalogram data.
2. Aiming at the difficult cross-tested situation in the field of electroencephalogram research, the invention carries out iterative optimization by combining a double-mapping-domain adaptive model and a semi-supervised regression and iconic notation propagation combined model and carries out projection matrix P aligning source domain data and target domain data in a shared subspace s ,P t A sample-label mapping matrix W, a sample-feature correlation matrix S and a target domain emotion label F of semi-supervised regression t Performing joint iterative optimization, and obtaining better alignment data by continuously iteratively optimizing the feature sharing subspace so as to be better applied to the learning of the sample-label mapping matrix W and the sample-feature incidence matrix S, thereby obtaining a more accurate target domain emotion label F t The method is applied to learning of the shared subspace.
3. The invention provides two submodels, namely a double mapping domain adaptation model and a semi-supervised regression and icon label propagation combined model. The double-mapping pre-adaptation model adopts a method of respectively mapping a source domain and a target domain to a shared subspace, and simultaneously combines the constraints of mapping matrixes of the two domains, thereby not only keeping the inter-domain common characteristics, but also not completely neglecting the respective beneficial individual characteristics of the two domains; a semi-supervised regression and icon label propagation combined model is adopted, a multi-model combined mode is combined with the semi-supervised regression model to jointly identify labels of a target domain, the semi-supervised regression model is used for depicting beneficial correlation in the domain, and the label propagation method based on a sample-feature bipartite graph is used for depicting label propagation between a source domain and the target domain based on features based on cognition that similar samples have similar feature distribution, so that possible adverse effects of sample correlation in the domain on a label propagation process are ignored, a conventional bipartite graph construction mode based on correlation between samples is abandoned, a feature level is added, and label propagation between the two domains is carried out from the perspective of the similarity degree of feature distribution.
Drawings
FIG. 1 is a frame diagram of an emotion recognition method for brain electrical migration;
FIG. 2 is a flow chart of an emotion recognition method for brain electrical migration;
FIG. 3 is a representation of a constructed sample-feature bipartite graph.
Detailed Description
The method identifies the sample of the unknown emotional state label by performing combined iteration on a double-mapping domain adaptive model and a semi-supervised regression and icon label propagation model, and for the double-mapping pre-adaptive model, a source domain and a target domain respectively correspond to a respective shared subspace mapping matrix P s ,P t Meanwhile, certain constraint is carried out on the difference possibly caused by the two domains in the mapping process; for the semi-supervised regression and graph semi-supervised model, the conventional bipartite graph is constructed from the angle of correlation between samples, specifically, when labels are transmitted, propagation is carried out through correlation between the samples, generally, distance between the samples, and it is considered that the same type of samples should have similar feature distribution, so that a bipartite graph representing sample-feature correlation is expected to be constructed for label transmission, and simultaneously, because the bipartite graph completely ignores sample correlation in a domain, in order to further compensate possible beneficial effects of the correlation of the domain itself, a multi-model combination mode is adopted to combine the semi-supervised regression model on the basis of the label transmission of the semi-supervised graph, and a combined label identification model is formed. The invention is further explained below with reference to the drawings;
as shown in fig. 1 and 2, the electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation specifically includes the following steps:
step 1, electroencephalogram data acquisition.
The emotion of a person does not appear very strong under daily conditions, in order to acquire strong emotion information, a certain induction needs to be performed on a tested person, 4 film segments with obvious emotion tendencies are selected in the embodiment, the film segments are respectively played to different tested persons at different times for watching, and the electroencephalogram data of the tested person are acquired as an original emotion electroencephalogram data set by connecting the electroencephalogram cap leads to corresponding brain areas while watching the films.
And 2, preprocessing data.
Sampling the electroencephalogram data acquired in the step 1, wherein the sampling rate is 200hz, and then filtering out noise and artifacts through a band-pass filter of 1 hz-75 hz, wherein the noise and the artifacts are respectively filtered in 5 frequency bands: delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (14-31Hz) and Gamma (31-50Hz), the Differential Entropy (DE) of which is calculated as the sample matrix X:
wherein σ is a standard deviation of the probability density function; μ is the expectation of the probability density function.
It can be seen that the differential entropy signature is essentially a logarithmic form of the power spectral density signature, i.e.
The preprocessing of the electroencephalogram signals aims to improve the signal-to-noise ratio, so that the preprocessing effect of data is improved, and interference is reduced.
Step 3, establishing a brain electrical migration emotion recognition model combining semi-supervised regression and icon label propagation, and uniformly integrating a double mapping domain adaptation model and a combined label recognition model: the former obtains a shared subspace, reduces the distribution difference between a source domain and a target domain, and the latter performs joint label identification based on the obtained alignment data to obtain an optimized target domain label for the former to better align the condition distribution of the two domains, so as to obtain a joint optimized target function, wherein the specific steps are as follows:
step 3.1, establishing a double mapping domain adaptive model:
wherein, the first and the second end of the pipe are connected with each other,
respectively mapping the source domain data and the target domain data to mapping matrixes in a shared subspace, wherein a superscript T represents transposition, p represents a shared subspace dimension, and d represents an original sample dimension;
to augment the source domain data tag, Y s One-hot type codes of source domain labels are adopted, and c is the category number;
in order to augment the target domain data tag,
representing a column vector with elements all 1, F t One-hot type coding for target domain tags, n s ,n t The number of samples of the source domain data and the target domain data respectively, n ═ n s +n t Representing the total number of samples;
for diagonal matrices, diagonal matrix N dom The kth diagonal element is the reciprocal of the number of kth source domain or target domain data samples, where k is 1,2, …, c; wherein
Is composed ofThe matrix of centers is a matrix of centers,
is a unit matrix which is formed by the following steps,
represents the square calculation of Frobenius norm, | | · | | calcualting 2,1 Represents the calculation of a 21 norm, | | P s -P t || 2,1 The row sparsity constraint of the difference value of the mapping matrix of the two domains sharing subspace can be realized, and the over-expression of the individual characteristics of the two domains in the mapping process can be prevented.
Step 3.2, establishing a semi-supervised regression and icon label propagation combined model:
s.t.G≥0,G1＝1,F≥0,F1＝1,
wherein the content of the first and second substances,
is a sample-to-label mapping matrix,
a matrix of sample labels is represented that,
mathematically representing X in semi-supervised regression T The average of the difference between PW and F,
representing a sample-feature bipartite graph correlation matrix,
representing a sample-feature correlation matrix, 0 n ,0 p Are respectively dimension of
The all-zero matrix of (a) is,
a graph association matrix representing an initialization constructed based on current shared subspace data,
is a sample-feature correlation matrix initialized based on current shared subspace data,
a matrix of labels representing the samples and the features,
is a graph Laplace matrix, a diagonal matrix
The diagonal elements are the rows S of the matrix, Tr (-) represents the trace of solving the matrix, and the formula (2)
Is a standard formula of semi-supervised regression,
is a row sparsity constraint on the matrix W,
representing a square calculation of a 21-norm,
the method is based on semi-supervised label propagation of the graph, firstly, an initialization graph is constructed based on the feature distribution of the shared subspace data, then, an optimal sample-feature bipartite graph which is expected to be constructed and is more suitable for the label propagation process is optimized through continuous iteration based on the constructed initialization graph, and the bipartite graph of the expected construction does not generate great difference with the feature distribution of the shared subspace data, namely
And carrying out label propagation on the target domain label based on the graph Laplacian matrix corresponding to the constructed optimal sample-feature bipartite graph, thereby obtaining the optimal target domain label.
Step 3.3, combining the two models constructed in the steps 3.1 and 3.2, and carrying out combined optimization according to an optimization model shown in a formula (11):
Step 4, according to the optimization model established in the step 3.3, aligning the projection matrix P of the source domain data and the target domain data in the shared subspace s ,P t A sample-label mapping matrix W of semi-supervised regression, a sample-feature bipartite graph correlation matrix S and a target domain emotion label F t And performing joint iterative optimization.
Step 4.1, initialize the data label F of the target domain t And a feature label F d Has a value of 0.25, and an initial shared subspace mapping matrix P is constructed by Principal Component Analysis (PCA) s ,P t 。
Step 4.2, fixing the target domain data label F t And a shared subspace mapping matrix P s ,P t And updating the sample-label mapping matrix W, wherein the objective function is as follows:
the target formula is an unconstrained optimization problem, and is calculated by adopting a Lagrangian function, and the following Lagrangian function is constructed:
wherein, the matrix
Is a diagonal matrix with the ith diagonal element of
eps is a fixed very small constant value,
is the ith row vector of the matrix, i is the row sequence number and j is the column sequence number.
The vector b is derived according to equation (12) and assigned zero, resulting in:
then, formula (14) is substituted into formula (13) for elimination treatment:
the matrix W is derived according to equation (15) and assigned zero to yield:
step 4.3, fixing the sample-label mapping matrix W and the target domain data label F t Updating the shared subspace mapping matrix P s ,P t The objective function is:
definition matrix
Matrix array
Is a diagonal element respectively as a matrix (P) s -P t ) The inverse of the 2 norm of each row vector, then:
||P s -P t || 2,1 ＝Tr(P T MP) (18)
defining:
the lagrangian function is constructed by substituting equations (14), (18), and (19) into equation (17):
the matrix P is derived according to equation (20) and assigned a value of zero to yield:
(XHX T ) -1 (T+M)P+P(αWW T )＝(XHX T ) -1 (XHFW T ), (21)
equation (21) is solved using the Siervests equation.
Step 4.4, fix the shared subspace mapping matrix P s ,P t Target domain data tag F t And a feature tag F d Updating the sample-feature bipartite graph correlation matrix S and calculating a new feature label F d The objective function is:
s.t.G≥0,G1＝1.
resolution of formula (22) yields:
s.t.G≥0,G1＝1,
solving the formula (23) by adopting a line-by-line solving mode:
wherein
Is the row vector of the sample label matrix F,
is a feature label matrix F d The row vector of (c), let:
rewrite equation (24) to:
solving the formula (26) to obtain the updated sample-feature correlation matrix S', updating the matrix D and the matrix L accordingly, and then comparing the λ Tr (Y) in the formula (22) T LY) using lagrange function to the feature tag matrix F d Updating:
F d ＝(D 2 ) -1 G T F. (27)
step 4.5, fixing the sample-label mapping matrix W and sharing the subspace mapping matrix P s ,P t And sample-feature bipartite graph correlation matrix S and feature labels F d Update target Domain data tag F t The objective function is:
s.t.F≥0,F1＝1.
disassembling and expanding matrix
F＝[Y s ；F t ]And
will be in formula (28)
To convert to:
in formula (28)
To convert to:
λ Tr (Y) in the formula (28) T LY) to:
constructing a Lagrangian function based on equation (28), equation (30), equation (31), and equation (32):
matrix F is aligned according to equation (33) t Derivative and assign a value of zero:
wherein:
based on KKT condition F t As Φ ═ 0, we obtain:
wherein
Represents the matrix Z s/t The negative values in (a) are all replaced by 0,
represents the matrix Z s/t All positive values in (1) are replaced by 0 and the absolute value is taken for negative values, i.e.
Step 4.6, repeating steps 4.2, 4.3, 4.4 and 4.5 for multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P s ,P t Sample-feature bipartite graph correlation matrix S and target domain data label F t Joint iterative optimization of (2).
And 5, inputting the sample matrix X obtained in the step 2 into the objective function subjected to iterative optimization in the step 4 to obtain a corresponding predicted value label, wherein the predicted value label is the emotional state of the testee corresponding to the sample at the acquisition moment. The predicted results of the method are shown in table 1 compared with the prior art:
TABLE 1
In the table, the method is compared with four conventional methods and the prior art mentioned in the background technology, the model identification precision is tested and compared, the subject represents a tested person, the '1- > 2' represents that a source domain is tested 1, a target domain is tested 2, the rest are analogized, the numerical values in the table represent the model identification precision, four decimal places are reserved for each group of experimental precision, six decimal places are reserved for the average precision of the method, and the identification precision of the method is obviously superior to that of the rest five methods.
Claims (6)
1. The electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation is characterized by comprising the following steps of: the method specifically comprises the following steps:
step 1, collecting electroencephalogram data of a plurality of testees under c different emotional states;
step 2, preprocessing and extracting characteristics of the electroencephalogram data acquired in the step 1, wherein each sample matrix X consists of electroencephalogram characteristics of a testee, and a label vector y is an emotion label corresponding to the electroencephalogram characteristics in the sample matrix X; selecting two sample matrixes X from different testees as source domain data X respectively s And target domain data X t ；
Step 3, constructing an electroencephalogram migration emotion recognition model combining semi-supervised regression and icon label propagation, wherein the electroencephalogram migration emotion recognition model comprises a double mapping domain adaptation model and a semi-supervised regression and icon label propagation combination model;
step 3.1, establishing a double mapping domain adaptive model:
wherein, dom is s, t,
respectively mapping matrixes for mapping source domain data and target domain data into a shared subspace, wherein a superscript T represents transposition, p represents a shared subspace dimension, and d represents an original sample dimension;
to augment the source domain data tag, Y s One-hot type codes of source domain labels are adopted, and c is the category number;
in order to augment the target domain data tag,
representing a column vector with elements all 1, F t One-hot type coding for target domain tags, n s ,n t The number of samples of the source domain data and the target domain data respectively, n ═ n s +n t Representing the total number of samples;
for diagonal matrices, diagonal matrix N dom The kth diagonal element is the reciprocal of the number of kth source domain or target domain data samples, where k is 1,2, …, c; wherein
Is a central matrix, and the central matrix is a central matrix,
is a unit matrix which is formed by the following steps,
represents the square calculation of Frobenius norm, | | P s -P t || 2,1 Mathematically, the row sparsity constraint of the difference value of the mapping matrix of the two domains sharing subspace is realized, and the significance of the row sparsity constraint is to prevent the over-embodying of the individual characteristics of the two domains in the mapping process, | | · | computationally 2,1 Represents the calculation of a 21 norm;
step 3.2, establishing a semi-supervised regression and icon label propagation combined model:
wherein the content of the first and second substances,
is a sample-to-label mapping matrix,
a matrix of sample labels is represented that,
represents X T The average of the difference between PW and F,
representing a sample-feature bipartite graph correlation matrix,
representing a sample-feature correlation matrix, 0 n ,0 p Are respectively dimension of
The all-zero matrix of (a) is,
a graph association matrix representing an initialization constructed based on current shared subspace data,
is a sample-feature correlation matrix initialized based on current shared subspace data,
a matrix of labels representing the samples and the features,
is a graph Laplace matrix, a diagonal matrix
The diagonal elements are the rows S of the matrix, Tr (-) represents the trace of solving the matrix, and the formula (2)
Is a standard formula of semi-supervised regression,
is a row sparsity constraint on the matrix W,
is a bipartite graph construction and label propagation process;
step 3.3, combining the two models constructed in the steps 3.1 and 3.2, integrating the double mapping domain adaptation model and the semi-supervised regression and icon label propagation combined model into a unified frame for combined optimization, and calculating a target domain emotion label F t The optimization model is as follows:
wherein, the alpha, the gamma, the beta and the lambda are all adjustable parameters in the optimization process;
step 4, initializing a target domain data label F t And a feature label F d Constructing and obtaining a subspace mapping matrix P of a source domain and a target domain by using a principal component analysis method s ,P t (ii) a And then according to the target function of the joint optimization obtained in the step 3, sequentially mapping the sample-label by fixing the other variables and updating the other variableMatrix W, subspace mapping matrix P s ,P t Sample-feature correlation matrix S and target domain data label F t Optimizing, repeating the optimization process for multiple times, and realizing combined iterative optimization;
step 5, substituting the result obtained after the combined iterative optimization in the step 4 into a formula (3), and then substituting the electroencephalogram data characteristic X with unknown emotional state into the electroencephalogram data characteristic X m As target domain data X t Inputting the data into a formula (3), and solving a corresponding target domain data label F t Obtaining the characteristics X of the brain electrical data m And correspondingly acquiring the emotional state at the moment.
2. The method for brain-electrical migration emotion recognition combining semi-supervised regression and icon label propagation as recited in claim 1, wherein: sampling the acquired electroencephalogram data at the frequency of 200Hz in the step 2, and then filtering noise and artifacts by passing the sampled data through a 1-75 Hz band-pass filter; and then dividing the sample matrix into five frequency bands of 1-4Hz, 4-8Hz, 8-14Hz, 14-31Hz and 31-50Hz, and respectively calculating the differential entropy under each frequency band as the electroencephalogram characteristic in the sample matrix X.
3. The method for recognizing brain electrical migration emotion combining semi-supervised regression and icon label propagation as recited in claim 1, wherein: the step 4 specifically comprises the following steps:
step 4.1, initialize the data label F of the target domain t And a feature label F d Has a value of
And constructing an initial shared subspace mapping matrix P in a dimension reduction mode s ,P t ；
Step 4.2, fixing the target domain data label F t And a shared subspace mapping matrix P s ,P t And updating the sample-label mapping matrix W, wherein the objective function is as follows:
solving a formula (4) to obtain an updated sample-label mapping matrix W';
step 4.3, fixing the sample-label mapping matrix W and the target domain data label F t Updating the shared subspace mapping matrix P s ,P t The objective function is:
solving the formula (5) to obtain the updated shared subspace mapping matrix P s ′,P t ′；
Step 4.4, fix the shared subspace mapping matrix P s ,P t Target domain data tag F t And a feature label F d Updating the sample-feature bipartite graph correlation matrix S and calculating a new feature label F d The objective function is:
solving the formula (6) to obtain an updated sample-feature bipartite graph incidence matrix S' and a feature label F d ′；
Step 4.5, fixing the sample-label mapping matrix W and sharing the subspace mapping matrix P s ,P t And sample-feature bipartite graph incidence matrix S and feature labels F d Update target Domain data tag F t The objective function is:
solving the formula (7) to obtain an updated target domain data label F t ′；
Step 4.6, repeating steps 4.2, 4.3, 4.4 and 4.5 for multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P s ,P t Sample-feature bipartite graph correlation matrix S and target domain data label F t Joint iterative optimization of (2).
4. The method for recognizing brain electrical migration emotion according to claim 3, wherein the method comprises the following steps: in steps 4.1, 4.3 and 4.4, an optimized sample-label mapping matrix W and a characteristic label F are solved through a Lagrange function d Target domain data tag F t The objective function of (1).
5. The method for recognizing brain electrical migration emotion according to claim 4, wherein the method comprises the following steps: in step 4.2, the optimized subspace mapping matrix P is solved through Lagrange function and the Sierviostes equation s ,P t 。
6. The method for recognizing brain electrical migration emotion according to claim 4, wherein the method comprises the following steps: and 4.3, simplifying and optimizing the objective function of the sample-characteristic incidence matrix S by a row-based solving method, and then solving.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210516213.4A CN114818822A (en) | 2022-05-12 | 2022-05-12 | Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210516213.4A CN114818822A (en) | 2022-05-12 | 2022-05-12 | Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114818822A true CN114818822A (en) | 2022-07-29 |
Family
ID=82513295
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210516213.4A Pending CN114818822A (en) | 2022-05-12 | 2022-05-12 | Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114818822A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117096070A (en) * | 2023-10-19 | 2023-11-21 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Semiconductor processing technology abnormality detection method based on field self-adaption |
-
2022
- 2022-05-12 CN CN202210516213.4A patent/CN114818822A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117096070A (en) * | 2023-10-19 | 2023-11-21 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Semiconductor processing technology abnormality detection method based on field self-adaption |
| CN117096070B (en) * | 2023-10-19 | 2024-01-05 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Semiconductor processing technology abnormality detection method based on field self-adaption |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Optimizing spatial patterns with sparse filter bands for motor-imagery based brain–computer interface | |
| Shoaib et al. | Signal processing with direct computations on compressively sensed data | |
| Majumdar et al. | Robust greedy deep dictionary learning for ECG arrhythmia classification | |
| CN111461176A (en) | Multi-mode fusion method, device, medium and equipment based on normalized mutual information | |
| CN113157094B (en) | Electroencephalogram emotion recognition method combining feature migration and graph semi-supervised label propagation | |
| CN112244873A (en) | Electroencephalogram time-space feature learning and emotion classification method based on hybrid neural network | |
| CN114533086B (en) | Motor imagery brain electrolysis code method based on airspace characteristic time-frequency transformation | |
| WO2021057423A1 (en) | Image processing method, image processing apparatus, and storage medium | |
| CN109598219B (en) | Adaptive electrode registration method for robust electromyography control | |
| Saboksayr et al. | EEG-based emotion classification using graph signal processing | |
| CN113010013A (en) | Wasserstein distance-based motor imagery electroencephalogram migration learning method | |
| Jiang et al. | Analytical comparison of two emotion classification models based on convolutional neural networks | |
| CN114818822A (en) | Electroencephalogram migration emotion recognition method combining semi-supervised regression and icon label propagation | |
| Gao et al. | EEG classification based on sparse representation and deep learning | |
| CN113378650B (en) | Emotion recognition method based on electroencephalogram source imaging and regularization common space mode | |
| US20230315203A1 (en) | Brain-Computer Interface Decoding Method and Apparatus Based on Point-Position Equivalent Augmentation | |
| Miao et al. | LMDA-Net: a lightweight multi-dimensional attention network for general EEG-based brain-computer interface paradigms and interpretability | |
| Taha et al. | A computationally efficient blind source extraction using idempotent transformation matrix | |
| Hughes et al. | Novel signal shape descriptors through wavelet transforms and dimensionality reduction | |
| CN110432899B (en) | Electroencephalogram signal identification method based on depth stacking support matrix machine | |
| CN114419657A (en) | ECG identity recognition method based on local segment sparse representation | |
| Bich et al. | Aggressively prunable MAM²-based Deep Neural Oracle for ECG acquisition by Compressed Sensing | |
| CN113177482A (en) | Cross-individual electroencephalogram signal classification method based on minimum category confusion | |
| Aiordachioaie | On Thermal Image Classification with ANN and Similarity Based Classifiers | |
| Kurian | Deep Convolution Neural Network-Based Classification and Diagnosis of Heart Disease using ElectroCardioGram (ECG) Images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |